A hybrid wavelet-machine learning model for qanat water flow prediction

In many parts of semiarid and arid regions, qanats are the leading supplier of water demand for agricultural and drinking usage. Qanat is an ancient collecting water system, and qanat water flow (QWF) varies in different seasons and decreases gradually by pumping groundwater wells. The present research utilized a set of supervised machine learning (ML) models to predict the QWF in the Chaghlondi Aquifer in Iran using monthly intervals of 14 years (2007-2021). The wavelet transform (WT) technique was also applied to enhance the QWF prediction quality of ML models for three lead months utilizing QWF, precipitation, evapotranspiration, temperature and GWL signal datasets as input. The five widely used ML models, i.e., artificial neural network (ANN), adaptive neuro-fuzzy inference system, group method of data handling (GMDH), gene expression programming and least square support vector machine, were applied and then compared with their hybrid wavelet models. To assess the performance of the models, the following four evaluation criteria were employed: correlation coefficient (R), Nash-Sutcliffe efficiency (NSE), root means squared error (RMSE) and mean absolute error (MAE). The outcomes showed that the hybrid-wavelet ML considerably improved the standalone model performance. The best QWF predictions for a one-month ahead QWF prediction were acquired from the WT-GMDH model results from input scenario 3 with RMSE, MAE, R and NSE equal to 14.46, 10.78, 0.93 and 0.85, respectively. In addition, the result of this study indicates that ML's performance was improved by using wavelet transform for two and three months ahead of QWF predictions.